paper-reviewerlisted

Conduct a formal peer review of a scientific paper Heath did NOT author — for a journal he is reviewing for, or as a pre-submission read for a collaborator. Six-phase workflow with a Coordinator (segments paper + routes), 6 parallel specialist subagents (Methods Auditor, Logic Checker, Presentation, Journal Fit, Adversarial Reader, Citation Verifier), a Synthesizer, and a Refiner that applies the structural rubric (blocking vs advisory tags, 3:1 minor:major cap, action-oriented minor comments). Output is a markdown review tuned to Heath's voice via the bundled voice.md. Bundle reference content lives at `Shared drives/Blackmon Lab/ Projects/paper-reviewer/` (agents/, references/, checklists/, voice.md). TRIGGER on: "review this paper", "peer review", "referee for <journal>", "what do I think of this manuscript", a directory of review materials, mention of revision/rebuttal/response-to-reviewers. Distinct from `pre_submission_review` (which is for Heath's OWN drafts heading to submission, not papers he is revi
coleoguy/tealc · ★ 0 · AI & Automation · score 70

Install: claude install-skill coleoguy/tealc

# Paper Reviewer (TEALC) Help Heath produce a thorough, honest peer review of a scientific paper. The review must reflect **Heath's** judgment in **Heath's** voice — the subagents are calibration tools to surface things he might miss and to enforce rigor against published LLM-review failure modes (hallucinated citations, generic boilerplate, sycophancy). They do not vote on the verdict. ## Why this design The published research on LLM-generated peer reviews shows three dominant failure modes, each with documented mitigations the workflow applies: | Failure mode | Citation | Mitigation in this workflow | |---|---|---| | Hallucinated citations and quotes (2.6% of accepted papers carry fabricated refs; LLM reviews are worse) | NeurIPS 2025 | **Citation Verifier** subagent: dedicated grounding pass — every quote verified in the source PDF | | Systematic sycophancy (AI scored papers higher in 53.4% of pairs vs human review) | Latona 2024 | **Adversarial Reader** subagent: explicit counterweight to the Methods Auditor's neutral-domain read | | Generic boilerplate that lacks paper-specific grounding (60% baseline for single-shot LLM reviews) | Liang 2024 / D'Arcy 2024 | **Aspect-decomposed multi-agent** with a Coordinator that segments the paper and routes specialists to relevant sections — MARG showed 2.2× more "good" comments under this structure | | Drift from Heath's voice / committee-vote feel | n/a (project-specific) | **Refiner** + voice-match pass apply the bundled `voic